{
  "id": 296626,
  "title": "Use Machine Learning to Predict Price, Then Calculate Target ",
  "url": "/competitions/g-research-crypto-forecasting/discussion/296626",
  "author_name": "A Day in the Sprawl",
  "post_date": "2021-12-22T16:07:44.282000",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hello,</p>\n<p>It seems logical to me, we are going to use time series analysis to predict price, perhaps through black box methods like neural networks. But then, we will use those outputs as inputs to a function that will calculate target via a known formula.   Is this the right way to approach the challenge? </p>\n<p>Thanks! </p>",
  "messages": [
    {
      "id": 1626547,
      "postDate": "2021-12-23T01:10:45.893Z",
      "content": "<p>It is probably best just to predict the target directly considering the target is impacted by the price changes of all coins.  Predicting those price changes for each coin separately and trying to combine them manually will not outperform a machine learning model doing it directly. </p>",
      "rawMarkdown": "It is probably best just to predict the target directly considering the target is impacted by the price changes of all coins.  Predicting those price changes for each coin separately and trying to combine them manually will not outperform a machine learning model doing it directly. ",
      "votes": 2
    },
    {
      "id": 1626254,
      "postDate": "2021-12-22T16:07:44.283Z",
      "content": "<p>Hello,</p>\n<p>It seems logical to me, we are going to use time series analysis to predict price, perhaps through black box methods like neural networks. But then, we will use those outputs as inputs to a function that will calculate target via a known formula.   Is this the right way to approach the challenge? </p>\n<p>Thanks! </p>",
      "rawMarkdown": "Hello,\n\nIt seems logical to me, we are going to use time series analysis to predict price, perhaps through black box methods like neural networks. But then, we will use those outputs as inputs to a function that will calculate target via a known formula.   Is this the right way to approach the challenge? \n\nThanks! ",
      "votes": 1
    },
    {
      "id": 1628222,
      "postDate": "2021-12-24T16:33:05.440Z",
      "content": "<p>I tried it did not quite work got 0.0175 score , it changes literally nothing except making the problem more complex at first we were predicting the rate of change in the price , and now predicting the numerator and denominator of that separately , that exactly all that changes ,</p>",
      "rawMarkdown": "I tried it did not quite work got 0.0175 score , it changes literally nothing except making the problem more complex at first we were predicting the rate of change in the price , and now predicting the numerator and denominator of that separately , that exactly all that changes ,"
    },
    {
      "id": 1628129,
      "postDate": "2021-12-24T14:41:14.283Z",
      "content": "<p>It seems  a reasonable approach to predict the log returns independently for each stock. It probably depends on your machine learning model if that would give  better or worse answers than predicting the target directly. (I am trying both methods)</p>",
      "rawMarkdown": "It seems  a reasonable approach to predict the log returns independently for each stock. It probably depends on your machine learning model if that would give  better or worse answers than predicting the target directly. (I am trying both methods)",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1626547,
      "author_name": "Kyle Peters",
      "author_url": "",
      "post_date": "2021-12-23T01:10:45.893000",
      "content": "<p>It is probably best just to predict the target directly considering the target is impacted by the price changes of all coins.  Predicting those price changes for each coin separately and trying to combine them manually will not outperform a machine learning model doing it directly. </p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 1628222,
      "author_name": "Sayantan Mazumdar",
      "author_url": "",
      "post_date": "2021-12-24T16:33:05.440000",
      "content": "<p>I tried it did not quite work got 0.0175 score , it changes literally nothing except making the problem more complex at first we were predicting the rate of change in the price , and now predicting the numerator and denominator of that separately , that exactly all that changes ,</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1628129,
      "author_name": "",
      "author_url": "",
      "post_date": "2021-12-24T14:41:14.283000",
      "content": "<p>It seems  a reasonable approach to predict the log returns independently for each stock. It probably depends on your machine learning model if that would give  better or worse answers than predicting the target directly. (I am trying both methods)</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1626547": "It is probably best just to predict the target directly considering the target is impacted by the price changes of all coins.  Predicting those price changes for each coin separately and trying to combine them manually will not outperform a machine learning model doing it directly. ",
    "1626254": "Hello,\n\nIt seems logical to me, we are going to use time series analysis to predict price, perhaps through black box methods like neural networks. But then, we will use those outputs as inputs to a function that will calculate target via a known formula.   Is this the right way to approach the challenge? \n\nThanks! ",
    "1628222": "I tried it did not quite work got 0.0175 score , it changes literally nothing except making the problem more complex at first we were predicting the rate of change in the price , and now predicting the numerator and denominator of that separately , that exactly all that changes ,",
    "1628129": "It seems  a reasonable approach to predict the log returns independently for each stock. It probably depends on your machine learning model if that would give  better or worse answers than predicting the target directly. (I am trying both methods)"
  }
}